[VLLMModel] Make the reasoning field configurable

[Kh4L]

Gym previously set both reasoning_content and reasoning on every outgoing assistant message. That's correct for vLLM, where each version ignores the key it doesn't know but fatal for servers that treat them as aliases of one field. Kimi-K3's Rust frontend answers 400 duplicate field 'reasoning', which killed every agentic rollout on turn 2. I confirmed the behaviour by probing the live server rather than inferring it: content alone, reasoning_content alone, and reasoning alone are all accepted; only the pair is rejected.

The fix adds reasoning_field: Literal["both","reasoning","reasoning_content"] = "both" to VLLMModelConfig. Default is unchanged, so nothing existing is affected. Five tests: parametrized across all three modes asserting exactly which keys reach the server, one pinning the default, one confirming pydantic rejects a bad value. 110/110 pass, ruff clean, DCO signed-off and SSH-signed.

If it's useful for the PR description, the failure signature reviewers would recognise is:

openai.InternalServerError: Error code: 500 - ... "Failed to deserialize the JSON body
into the target type: messages[2]: duplicate field reasoning"

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.